Accidental fall seriously threatens the health and safety of the elderly. Accurately identify the behavior of human falls and giving timely alerts are effective means to reduce the damage of accidental fall-wound. In this paper，we present a new fall detection method. In our method，dynamic characteristics of human tilt posture are extracted from the key points of the human body based on OpenPose deep convolutional network，the dynamic characteristics are then used for Linear SVM to detect falls，a judgment based on human descending posture is made to exclude confusing human behavior and improve the recall rate. Our method has achieved 97.33% accuracy and 94.80% precision on the human motion dataset，which is better than the current image-based falling behavior recognition method. Being suitable for monocular RGB camera make our method superior in practicality to the existing falling behavior recognition methods that require Kinect cameras.